Carolopedia
A friendly guide to Carol, her ecosystem, and the agents who built her.
📖About & Usage
About
Judging the initiative — aggregating all gate signals into a stance (continue / replan / close / escalate) and running the close decision and its post-commit hooks.
Where it fits
This is one stage of the Build Initiatives service. The owner and the agents who run it are listed under the team below, and the other blocks of the service are linked at the bottom of this page.
🛠️Team & droids
Elrond Block owner
Judging is where all the gate signals collapse into one decision about the initiative's fate, and Elrond owns that decision end to end. The Elrond Supervisor is the aggregator: it reads the budget-gate status and the latest review verdict and gaps, counts how many times Merlin has already replanned, gathers open handshake requests from executors, then calls Claude to judge whether the gaps are serious enough to keep investing or minor enough to close — and applies Elrond's decision matrix to dispatch one of continue, replan, close, or escalate, executing that choice (API close, handshake to Merlin, notify Orion, or log continuation). The Initiative Review Inferer is the upstream trigger that flips an initiative to reviewing: whenever a plan step is marked done it re-reads the committed status, counts the steps in that phase, and if every step is done and no review row yet exists for the phase, flips the status to reviewing (it infers the transition only — it never reviews anything itself). Two Initiative Closer droids handle the terminal act: one closes on operator instruction, refusing if must-have criteria are unmet unless an override is passed; the other runs the four post-commit close-event hooks once a status change is committed — flipping the bypass dispatch lock on terminal status, rolling component actuals into running averages on close, running quality scoring plus the rolling-average regression check, and running the per-initiative audit plus follow-on filer on any terminal state. This matters because it is the only place that decides whether work continues, loops, ends, or goes to a human, and it makes that call from aggregated evidence rather than gut feel. It fires once steps are done and the reviewing flip has been inferred, with the close hooks firing on commit and the escalate path covering the genuinely ambiguous cases.